University Defence Research Collaboration (UDRC) Signal

University Defence Research Collaboration (UDRC)Signal Processing in a Networked Battlespace

WP5: Networked enabled sensor managementWP Leader: D. Clark, Academics: Y. Petillot, M. Davies

Researcher: E. Delande

Introduction: The aim of this work package is to investigate the resolution ofsensor management problems involving hierarchical information-based decisions at dif-ferent levels of the signal processing chain (sensor modes and multi-sensor controlpolicies). The general goal is to provide a unified framework for multi-object Bayesianfiltering (i.e. multi-target detection and tracking), multi-sensor data fusion and sensormanagement. The Finite Set Statistics framework, developed for the construction ofvarious multi-sensor/multi-target tracking filters following a principled approach, isflexible enough to incorporate sensor management policies into a unified frameworkand will be the backbone of the developments in this work package.

Objective:

WP 5.1: Hierarchical sensor management to target tracking (→ WP2, WP3)

• Unify multi-object Bayesian estimation, multi-sensor data fusion, and sensormanagement;

• Focus on novelty and clarity of proposed solutions.

WP 5.2: Computationally tractable solution (→ WP6)

• Improve the implementation and reduce the computational costs of the newalgorithms;

• Focus on parallel processing, SMC/GM implementations, optimal allocationproblems.

WP 5.3: Multi-objective sensor management

• Explore additional objectives beyond target detection/tracking;

• Focus on prioritisation of multiple objectives.

Multi-object Bayesian estimation

How many targets? Where are they?

http://www.nollywoodone.com/latest-additions/9009-the-u-s-military-s-real-time-google-street-view-

airborne-spy-camera-can-track-an-entire-city-in-1-

800mp.html

Stochastic population of objects modelled withpoint processes or Random Finite Sets:

• Target population;

• Target interactions;

• Target measurements;

• Missed detections;

• False alarms;

• Target appearing/disappearing;

• ...

Multi-object filters may propagate:

• Regional information on targetactivity (i.e. mean target numberµΦ(B));

• Individual information on specifictargets (probability of existence,spatial distribution).

Estimation of regional activity

Figure 1: Variance in target number:PHD filter (top), CPHD filter (bottom)

Regional information statistics:

• Retrieved from functional repre-sentation of point processes;

• Similar to moments for randomvariables;

• 1st order: mean target numberµΦ;

• 2nd order: variance in targetnumber varΦ.

Statistics (µΦ(B), varΦ(B)) provide:

• An estimation of target activity inB;

• An associated uncertainty.

Research axis: exploit (µΦ(B), varΦ(B))for sensor-level management decision (→WP2).

Multi-sensor fusion: comparative study

Goal: assess multi-sensor multi-objectBayesian filters

Assessment tool: regional statistics(µΦ, varΦ)

Research axis: propose a reference filterfor future performance assessment

Performance evaluation: reference filter

The multi-target Bayes filter is optimal, but intractable in the general case. What ifthe targets are assumed independent?

Research axis:

1. Design and implement a reference filter for independent targets;

2. Produce and exploit its regional mean and variance to assess the available fil-ters (“CRLB for multi-object filters”).

Ongoing development: the Independent Stochastic Process (ISP) filter

• Design completed (→ collaboration with J. Houssineau), implementation on-going;

• Individual tracks are identified by an observation path (i.e. an history of mea-surements);

• Regional statistics are easily extracted from the filter output;

• Computationally intensive, room for efficient implementation (→ WP6).

Application: modelling of Doppler radar

(a) SNR = 9dB, Pfa = 10−3, Pd =0.65

(b) SNR = 9dB, Pfa = 10−4, Pd =0.43

Modelling of Doppler radar:

• Parametrisable by physical characteristics (pulse, bandwidth, etc.);

• Stochastic description (SNR, Pfa, Pd) of local sensor behaviour in an elemen-tary cell → well adapted to ISP filter;

• Two degrees of freedom among SNR, Pfa, and Pd.

Research axis: integrate stochastic description of sonar systems (→ WP3).

Conclusion:Recent developments in multi-object Bayesian estimation techniques allow:

• The design of multi-object filters propagating information on target populationas well as individual targets;

• The construction and exploitation of higher-order regional statistics estimatingthe level of target activity anywhere in the surveillance scene.

Future Work:1. Design of an assessment tool for multi-object solutions based on the ISP filter;

2. Exploitation of regional statistics for sensor-level control policies;

3. Integration of multi-sensor management to multi-object Bayesian estimationframework.

This work was supported by the Engineering and Physical Sciences Research Council (EPSRC) and Dstl.